[Data] effect of virtual_loss

[sethtroisi]

I'm playing games with various vloss, this first set was played with 400 playouts. I'm going to attempt a 2nd set with 800, and then play some with even time.

minigo-v10-000593-1 v minigo-v10-000593-2 (64/64 games)
board size: 19   komi: 7.5
                      wins              black         white       avg cpu
minigo-v10-000593-1     33 51.56%       16 50.00%     17 53.12%    218.25
minigo-v10-000593-2     31 48.44%       15 46.88%     16 50.00%    164.80
                                        31 48.44%     33 51.56%

minigo-v10-000593-1 v minigo-v10-000593-4 (64/64 games)
board size: 19   komi: 7.5
                      wins              black         white       avg cpu
minigo-v10-000593-1     39 60.94%       20 62.50%     19 59.38%    214.17
minigo-v10-000593-4     25 39.06%       13 40.62%     12 37.50%    113.68
                                        33 51.56%     31 48.44%

minigo-v10-000593-1 v minigo-v10-000593-8 (64/64 games)
board size: 19   komi: 7.5
                      wins              black         white       avg cpu
minigo-v10-000593-1     42 65.62%       22 68.75%     20 62.50%    205.69
minigo-v10-000593-8     22 34.38%       12 37.50%     10 31.25%     94.62
                                        34 53.12%     30 46.88%

minigo-v10-000593-2 v minigo-v10-000593-8 (64/64 games)
board size: 19   komi: 7.5
                      wins              black         white       avg cpu
minigo-v10-000593-2     41 64.06%       21 65.62%     20 62.50%    153.73
minigo-v10-000593-8     23 35.94%       12 37.50%     11 34.38%     93.77
                                        33 51.56%     31 48.44%

[sethtroisi]

The main takeaway for me here is that batch_size_2 ~= batch_size_1 but runs in 33% faster (or in 75% of the time).

it's possible that batch_size_4 ~= batch_size_2 when playouts is 800. I might explore that at some point.

[sethtroisi]

We are testing the effect of batch_size (AKA virtual losses) in an RL run (v12), be reducing BS=8 (v10) to BS=2 (v12) but holding other variables constant, see details on why this might be important in #443

Currently at ~120k steps the runs look identical which is reassuring about our pipeline, reproduciblity, and the speedup of BS=8 but someone disappointing about a potential for a stronger network.

[amj]

@sethtroisi have we got playout strength of 1 vs 8 vs 128?